GLM 5.3 Flash on 4x 8 GB/Confirmation 7
From CMP Wiki
More actions
| Confirms | GLM 5.3 Flash on 4x 8 GB |
|---|---|
| Result | Reproduced |
| By | @con3po7016 |
| Date | 2026-09-26 |
| Cards | 8 GB |
| Numbers | Release 1.4.0 on 4 cards at Gen2 x8, 200 W per card. TP4 INT4: one user 271 / 262 / 184 tok/s (structured / code / prose), eight users 756 / 733 / 495, prefill 1.9K / 2.0K at 16K / 64K. PP4 MM INT4: 144 / 136 / 99, eight users 650 / 647 / 423, prefill 4.8K / 7.4K. PP4 KV pool 2,320,328 tokens, reproduced exactly from the x16 figure. GSM8K 97.2 / 98.8, MMLU (300) 91.7, tools 24/24. |
Four of the eight cards in his box, measured against his own older 6-card engine. On x8 links PP4 matches the x16 numbers and TP4 decode reaches 90 to 98% of them; TP4 prefill is about 62% (1.9K against 3.1K) because x8 halves the prefill all-reduce bandwidth. The engine's outputs were bit-deterministic run to run. His configuration files are not reproduced (asked of the author). Source: @con3po7016, 2026-09-26